Adaptive feature-extraction graph network for physical systems: Prediction of inviscid compressible flow in urban explosion

Abstract

High-fidelity numerical simulation of complex physical systems (e.g., urban explosion) is computationally demanding owing to the precise modeling of the intricate physical interactions within large spatial domains. Machine learning-based surrogate models provide high efficiency but are often limited by direct input-output mappings that fail to capture underlying physical laws, reducing generalizability. Graph neural networks (GNNs) offer a potential solution, but often exhibit sensitivity to input features and limited adaptability across systems with diverse inputs. In this study, we propose an Adaptive feature-extraction graph network (AdaFGN), which integrates an adaptive feature extractor with the advanced GNN modules. The extractor learns informative feature representations, while GNN modules model the underlying dynamic physical interactions. Validation results on compressible flows, including air blasts and urban explosions, demonstrate that AdaFGN: 1) predicts pressure fields with an average RMSE of 0.009; 2) generalizes robustly to unseen and real-world scenarios; 3) offers effective and flexible feature engineering; and 4) maintains efficiency with only a 2.2 % increase in inference time compared to GNNs without the feature extractor. These advantages stem from effective feature extraction and physical interaction modeling, establishing AdaFGN as a robust surrogate model for physical simulation.